Snap AI Coding Interview, Experienced SWE and MLE: 4 Experiences

Shared by janebeingjane ·
Company
Snap Inc.
Reports
4 interview reports
Level
Full-time
Stages
Onsite

Snap's 2026 onsite AI coding round had software engineers build a backend chat service from an empty Python project with Claude or Codex.

At a glance

  • Process: one AI coding round per onsite, next to system design and regular coding rounds (01–03).
  • Asked recently: the same backend-only chat service in 02 (L5) and 03 (L4); a feature in an existing codebase in 01. The ML engineer's "AI coding" round had no AI in it (04).
  • Difficulty: candidates planned first: 01 had the AI explain the codebase, and 03 wrote a plan document with the interviewer before implementing.
  • Tools: 03 could pick Claude or Codex.

Coverage: Four experienced full-time hires, all changing jobs: three software engineers (one at L4, one at L5, one on an ML infrastructure team with no level given) and one machine learning engineer. All four describe the onsite; two of the software engineers also mention the phone screen. The interviews took place between July and September 2026, and the reports went up between August 12 and November 12, 2026. No locations were given.

Roles in this article

  • Software engineer, 3 reports: one AI coding round per onsite (a chat backend twice, a feature in an existing codebase once), plus system design and ordinary coding rounds.
  • Machine learning engineer, 1 report: a coding round billed as AI coding that was really a classic graph question, plus recommender-system design rounds.

Software engineer: Snap AI coding onsite (3 reports)

What came up

Every software engineer here had exactly one AI coding round at the onsite, sitting next to system design and regular coding rounds. Two of them got the same assignment, a chat backend, and the one who described the tools could pick Claude or Codex.

Topic Reports Last seen
One AI coding round at the onsite 3 of 3 Sep 2026
System design rounds in the same loop 3 of 3 Sep 2026
Regular coding rounds in the same loop 3 of 3 Sep 2026
AI agent named: Claude Code, or a choice of Claude or Codex 2 of 3 Sep 2026
Building a chat backend from an empty Python project 2 of 3 Jul–Sep 2026
Adding a feature to an existing codebase 1 of 3 Sep 2026

A report can count toward several rows.

Questions we found

AI Coding

Question Reports Last seen
Build a Persistent Chat Backend: backend-only chat in an empty Python project with send and receive, offline delivery and a conversation log, then a move from memory to a database 2: 02, 03 Jul–Sep 2026
Add a feature to an existing app with a Python backend and a JavaScript frontend (the feature wasn't named; the change landed in the frontend) 1: 01 Sep 2026 (report date)

Algorithms / Data Structures

Question Reports Last seen
Number of Closed Islands 1: 01 Sep 2026 (report date)
Radar Barrier Crossing: can something cross a rectangle without touching any radar circle 1: 01 Sep 2026 (report date)
Stream Cluster Max and Median (phone screen) 1: 01 2026, before a Sep 2026 onsite
Sort Characters by Frequency 1: 02 Jul–Sep 2026
Dictionary Records to an Editable XML Tree 1: 03 Jul–Sep 2026
Metric Counter with a Fake Clock 1: 03 Jul–Sep 2026

System Design

Question Reports Last seen
CI/CD system with a control plane and data plane, spreading worker hosts across regions (CI/CD Pipeline Orchestration System) 1: 01 Sep 2026 (report date)
Metrics system (Design a Time Series Metrics Database) 1: 02 Jul–Sep 2026
Pre-login risk challenge service (Design a Pre-Login Risk Challenge Service) 1: 02 Jul–Sep 2026
WhatsApp (One-to-One Chat Service) 1: 03 Jul–Sep 2026

3 detailed candidate timelines

01: ML infra engineer whose AI round switched from Go to Python on the spot

Report date: Oct 7, 2026 · Interview date: September 2026 · Role: Software engineer, ML infrastructure, level not given · Experience: changing jobs, applied online · Result: Rejected

Phone screen

  1. Data arrives as a stream. Each new value either joins one of the stored clusters or starts a new one, and you need a way to read each cluster's maximum. Follow-ups asked how to make reads and writes faster and how to get each cluster's median. Practice version: Stream Cluster Max and Median.

Onsite: ML infrastructure system design

  1. The candidate had prepared for distributed training across several orgs and feature pipelines. The actual question was a CI/CD system with a clear split between the control plane and the data plane, and worker hosts spread out by region. Practice version: CI/CD Pipeline Orchestration System.

Onsite: coding 1

  1. Number of Closed Islands.
  2. The candidate thought the interviewer was pasting the code into an AI in the background to check it. The interviewer then came back with two test cases that failed.

Onsite: coding 2

  1. Several radars, each a circle with a center and radius. Find which radars overlap and group them with union-find, then decide whether something can cross a rectangle from the left wall to the right wall without being detected. It can't if one group of radars covers the rectangle from top to bottom. Practice version: Radar Barrier Crossing.

Onsite: AI coding

  1. The prep meeting and email said this round would be building a small program from scratch, covering both design and the full development cycle.
  2. On the day it was different. The interviewer said it was only their second time running this round, and the task was in an existing codebase.
  3. That codebase was in Go. The candidate asked for Python, C++ or Java, and the interviewer found a "Python" task on the spot: a Python backend with a JavaScript frontend.
  4. The candidate used AI to understand the architecture first and found that the feature only needed a frontend change. They then built it by working with Claude Code.
  5. The report doesn't say what the feature was, so this task has no practice version.

02: L5 loop with the same chat backend task

Report date: Sep 10, 2026 · Interview period: Jul–Sep 2026 · Role: Software engineer, L5 · Experience: changing jobs · Result: Rejected two days after the onsite

Onsite: behavioral and system design 1

  1. Behavioral and résumé questions first, then about 30 minutes of system design: a metrics system. No scale or query details were given. Practice version: Design a Time Series Metrics Database.

Onsite: coding

  1. Sort Characters by Frequency.

Onsite: system design 2

  1. A pre-login risk challenge system. Before a user logs in, the service looks at their information, judges the risk and returns a list of challenges they must pass, CAPTCHA being one of them. There's also an API to add a new challenge. Practice version: Design a Pre-Login Risk Challenge Service.

Onsite: AI coding

  1. A message backend service. The candidate said it was the same task as in report 03 below, which matches the practice version Build a Persistent Chat Backend.
  2. Nothing was said about the tools or how far they got.

03: L4 loop: a chat backend with Claude or Codex

Report date: Aug 12, 2026 · Interview period: Jul–Sep 2026 · Role: Software engineer, L4 · Experience: changing jobs, applied online · Result: Not given

Phone screen

  1. Two easy coding questions built around loops, plus a few basic behavioral questions. The questions weren't named.

Onsite: AI coding

  1. The starting point was an empty Python backend. The candidate could use Claude or Codex to build a chat system, backend only, with no UI.
  2. They spent a long time clarifying the requirements with the interviewer and settled on three features: sending and receiving messages, getting messages that arrived while the user was offline, and viewing a conversation's history.
  3. They wrote a plan document first, then implemented it.
  4. Once the basics worked, they moved storage from memory to a database.
  5. The interviewer then asked feature-level questions: polling versus push for delivering messages, and how to handle unread messages.
  6. The candidate's view: not hard if you already code with AI at work.
  7. Practice version: Build a Persistent Chat Backend covers send and receive, offline delivery, history and database storage. The polling-versus-push and unread-message follow-ups were discussion only.

Onsite: coding 1

  1. Given a list of dictionaries, rebuild the XML they describe, then support adding and removing elements. Practice version: Dictionary Records to an Editable XML Tree.

Onsite: coding 2

  1. A metric counter, with unit tests that use a fake timer. Practice version: Metric Counter with a Fake Clock.

Onsite: system design

  1. Design WhatsApp. Practice version: One-to-One Chat Service.

Machine learning engineer: Snap onsite with an "AI coding" round (1 report)

What came up

The one MLE report had a coding round that was called AI coding, but the candidate says there was effectively no AI part: it was a classic graph question. The other rounds were two recommender-system design rounds and an ML fundamentals deep dive on the résumé.

Questions we found

Algorithms / Data Structures

Question Reports Last seen
Graph traversal over parent-child relations with fairly involved conditions, in the round billed as AI coding 1: 04 Jul–Sep 2026

ML System Design

Question Reports Last seen
Recommender system, one round on retrieval and one on ranking, with ads questions such as how labels are computed (Design a Personalized Recommendation System, which has no ads part) 1: 04 Jul–Sep 2026

1 detailed candidate timeline

04: MLE virtual onsite where the AI coding round was plain coding

Report date: Nov 12, 2026 · Interview period: Jul–Sep 2026 · Role: Machine learning engineer · Experience: changing jobs, came through a headhunter · Result: Rejected

Onsite: coding (billed as AI coding)

  1. The round carried the AI coding name, but the candidate says it might as well not have: it was a traditional coding question.
  2. The task gave some parent-child relations and asked for a graph traversal under fairly involved conditions. The candidate ran a BFS over one level, then a DFS. The conditions and expected output weren't shared, so there's no practice version.

Onsite: ML design and applied ML

  1. Both rounds were about recommender systems. On paper one leans toward system details and the other toward ML modeling, but to the candidate they felt about the same, with a bit of everything asked.
  2. One focused on retrieval, the other on ranking. One of them went into ads details, such as how labels are computed.
  3. Closest practice version: Design a Personalized Recommendation System, which covers retrieval and ranking but not ads.

Onsite: ML fundamentals

  1. A deep dive that followed the résumé. The candidate rated it the best round of the loop, with many interesting detailed questions.

This report also has its own page: 2026 Snap Experienced Machine Learning Engineer Virtual Onsite.


Patterns across the reports

These come from the three software engineer reports (01–03). Report 04 is a different role and stays out of the comparison.

  1. One AI round per loop, among regular rounds. Each onsite had a single AI coding round. The rest was system design and ordinary coding: in 01, one design and two coding rounds; in 02, two design rounds and one coding round; in 03, two coding rounds and one design round.

  2. The chat backend came up twice. Reports 02 (L5) and 03 (L4) got the same message backend task, and the L5 candidate pointed to the L4 account as the same question. Report 01, interviewed in September, got a feature in an existing codebase instead, from an interviewer who said they had only run that round once before.

  3. Understanding or planning before the agent writes code. In 01, the candidate first had the AI explain the codebase's architecture, which showed the change only touched the frontend. In 03, the candidate pinned down requirements with the interviewer and wrote a plan document before implementing.

Common questions

What does Snap ask in the AI coding interview?

Two of three software engineers built a backend-only chat service from an empty Python project (02, 03); 01 added a feature to an existing codebase.

Does Snap let you choose the AI model?

Candidate 03 could choose Claude or Codex; the others didn't say.

Keep practicing

AI coding rounds at other companies

More companies with AI rounds on FastPrep:

Across companies: 2026 AI Coding OAs at 15 Companies: 32 Reported Problems and 2025–2026 AI Coding Phone Screens and Onsites at 43 Companies: 76 Reported Problems.

Sources: Anywhere useful interview info can be found.

0 total reactions. Your reactions: none.

Comments

0

No comments yet — be the first to share what you know.